DeepSeek V2: Complete Benchmark Performance Guide

deepseek v2 2024-05-07 DeepSeek License

In-depth analysis of DeepSeek V2 performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendordeepseek
Versionv2
Release Date2024-05-07
Context Window32768 tokens
Input Modalitiestext
Output Modalitiestext
LicenseDeepSeek License
Documentationhttps://api-docs.deepseek.com/

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 78.2 % 2024-05-07 5-shot view
HUMANEVAL 75.7 pass@1 2024-05-07 — view
GSM8K 78.8 % 2024-05-07 0-shot CoT view
MATH 35.2 % 2024-05-07 0-shot CoT view
BBH 70.5 % 2024-05-07 3-shot CoT view
GPQA 31.5 % 2024-05-07 0-shot view
IFEVAL 78.6 % 2024-05-07 prompt_strict view
ARC 92.1 % 2024-05-07 challenge view
MUSR 53.4 % 2024-05-07 0-shot view
WINOGRANDE 84.7 % 2024-05-07 0-shot view

Pricing

Tier Price Currency
Input$0.14 / MtokUSD
Output$0.28 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://api-docs.deepseek.com/quick_start/pricing · as of 2024-05-07

Compliance

  • Data Residency: CN
  • SOC2: ✗
  • HIPAA: ✗
  • GDPR: ✗
  • ISO 27001: ✗

DeepSeek V2

Model Overview

DeepSeek V2 开源 MoE 模型, 总参 236B/活跃 21B, 32K 上下文, 首创 MLA 注意力机制, 性价比突出。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Deepseek v2 2024-05-07 32K text text DeepSeek License

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 78.2 % 5-shot
HumanEval 75.7 pass@1 —
GSM8K (Grade School Math 8K) 78.8 % 0-shot CoT
MATH 35.2 % 0-shot CoT
BBH (BIG-Bench Hard) 70.5 % 3-shot CoT
GPQA 31.5 % 0-shot
IFEval 78.6 % prompt_strict
ARC 92.1 % challenge
MUSR 53.4 % 0-shot
WinoGrande 84.7 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Mixture-of-Experts architecture.

Weaknesses

  • Proprietary, not self-hostable.

Use Cases

  • Code generation and debugging

References